Counting processes for spatial capture-recapture

Abstract

Estimates of animal density, the number of individuals per unit area, are critically important for understanding ecological processes affecting wildlife management. Increasingly, modern technology like camera traps and acoustic recording units are being used to monitor wildlife populations. These are relatively inexpensive and can reliably record animal detections over a long period of time. When animals have a chance to be recorded on multiple detectors, spatial capture-recapture (SCR) can be used to estimate animal density. However, these models generally require that animal identity is known which is often not the case. For camera trap studies, the animals may not have unique pelage to distinguish them reliably, or image quality may be poor. In an acoustic recording, there may not be any individually identifying features in the individual’s call to distinguish it from others of the same species. The motivating problem of this thesis is how to deal with uncertain animal identity in SCR. We review current methods in SCR for both known and unknown identity, ways to make Bayesian inference for SCR models using Nimble in R, and how SCR can be written as a Dirichlet process mixture model when identities are unknown. This leads us to reformulate the conventional SCR model as a marked Poisson process, such that the counting process for detections through time no longer depends on identity, but the observed mark distributions do. When identity is latent, the observed marks are distributed over a mixture of N latent animal characteristics (e.g. activity centre, sex), where N is the number of animals at risk of detection. This becomes a generalization of the unmarked SCR model of Chandler and Royle (2013) and allows us to easily add additionally observed covariates to help estimate animal identity. We show through simulation how well the method works and apply it to a camera trap survey of fisher (Pekania pennanti) and an acoustic survey of the Cape Peninsula moss frog (Arthroleptella lightfooti), each with different types of information used to inform identities. A fundamental assumption of SCR is that detections of an individual occurs independently. This implies that a detection at one detector has no impact on the probability that an animal is seen at any other detectors immediately after. Assumptions of independence allows us to model SCR detections as a spatiotemporal point process, often a Poisson process. As a result, the time of detection becomes uninformative about animal identity. We offer a solution to this by thinking about SCR in terms of a new detection function that depends on a realistic animal movement model. To do this, we relax the independence assumption by building a spatiotemporal dependent point process with a new detection function that depends on where and when the animal was last observed. As a result, we can explicitly model the existing correlation in the detections as well as provide a new method for spatiotemporal clustering of latent identity SCR problems.